Combining wavelets and watersheds for robust multiscale image segmentation

This paper proposes a new segmentation technique that combines multiresolution wavelet decompositions with the watershed transform. The wavelet transform is applied to the intensity image, producing detail and approximation coefficients. Gradient magnitudes of the approximation image at the coarsest resolution are computed, and an adaptive threshold is used to remove small gradient magnitudes. The watershed transform is then applied, and the segmented image is projected up to higher resolutions using inverse wavelet transforms. Typically, if a low resolution is chosen for the initial segmentation, large relevant objects will be captured; on the other hand, a higher initial resolution will lead to smaller (and more detailed) segmented objects. The low-pass filtering involved in the wavelet decomposition provides robust segmentation results for noisy images, even when the amount of noise is very large.